Went through the Lyft L5 loop in March. System design is one round, 45 minutes, and it's more scoped than what you see at Google or Meta. They're not going to ask you to design YouTube. Expect something closer to their actual domain: ride matching, pricing systems, location tracking, notifications at scale, driver-rider pairing under latency constraints.
The round I had: design a real-time surge pricing system. Not a toy problem. They cared about: how you decompose the write path vs. read path, what your data model looks like at the driver supply level, how you handle partition tolerance when a region's location service goes down.
What they actually care about: Can you scope the problem yourself (they will not do it for you) Do you reason about tradeoffs out loud or just state your preferred answer Do you know the difference between a queue-based fanout and a pull model and when you'd pick each Failure modes: what breaks first and how you detect it
I went pretty deep on Kafka for the event pipeline and they pushed back to ask what happens if you have a consumer group lag spike during peak hours. Good question, caught me slightly off guard. Have an answer for graceful degradation.
They also ask about observability more than most companies. What metrics would you alert on, how would you know the system is degrading before a customer files a complaint. Worth prepping a short answer on SLOs and how you'd instrument the critical path.
Level-wise: at L5 they expect you to drive the session, not wait for prompts. If you're sitting back waiting to be asked the next question, that reads junior. Walk them through your thinking, checkpoint with them when you're about to go deep somewhere, and ask explicitly if you should.
Total loop was 5 rounds: 1 system design, 2 coding, 1 behavioral, 1 leadership/cross-functional. System design was the round people seemed most anxious about but honestly it's the most straightforward if you do real systems in your day job.